clef-NVFP4 / scripts

Commit History

  1. 8-bit lookup tables: 20.0 -> 17.5 GB, same accuracy and speed; faster server (prompt-sized CUDA graphs, --gdn-backend)

    Verified signature
    1abd566

  1. Upload clef-NVFP4 (Clef 27B NVFP4, vLLM plugin + SystemOne server)

    Verified signature
    6ba50d1